Estimating Fibrosity Scores of Plant-Based Meat Products from Images: A Deep Neural Network Approach

Automation of the plant-based meat extrusion process requires a scheme to provide quantitative estimates of output fibrosity, which must be carried out online and in real-time. A novel machine learning regression model for this purpose, is proposed in this article. A deep neural network that was originally trained for image classification, was extended to provide quantitative fibrosity estimates. Relevant layers of the network were retrained using real-world laboratory data. Plant-based meat or textured vegetable protein products with varying fibrous microstructures were obtained using different ingredient formulations and process conditions on a pilot-scale twin screw extruder with in-barrel moisture range of 29.2-40.9% (wet basis). Images of extruded plant-based meat products were collected to serve as sample inputs. An experiment was devised, where image samples were randomly presented to two expert human subjects, who provided as feedback, fibrosity scores lying within the interval [1, 10]. Statistical metrics were adopted to evaluate the performance of the trained network. It was found that the network performed significantly better when trained separately with feedback scores of each individual subject, than with the combined scores, indicating that it was able to capture nuances of a subject’s perception. Another study was directed at the explainability of the network’s estimations. Using standard software, a set of synthetic images of varying shapes and sizes were created as inputs to the network. Interpretations of its output scores indicate that the network’s estimates were based on features relevant to porosity and fibrosity, while not influenced by extraneous ones.